NVIDIA's AVO harness lifts Opus 5 from 30% to 100% on ARC-AGI-3
daniel_mac8 · x · 2026-08-22
NVIDIA solved the ARC-AGI-3 public benchmark with its agent harness AVO + Opus 5, boosting Opus 5 from 30% under the standard harness to 100%. The poster breaks down what the harness does to expose an LLM's latent capabilities:
- Persistent memory across context windows
- An inspect → plan → implement → evaluate loop
- Execution feedback so the agent learns from mistakes at inference time
- A supervisor that detects stagnation in failed trajectories and redirects
- External statefulness of the transcript (tool calls, execution artifacts, agent trajectory)
His takeaway: frontier harness engineering doesn't require working at a frontier lab — harness advances don't need to touch model weights but can make all the difference.
Related event: NVIDIA's AVO Agent Hits 100% on ARC-AGI-3 Public Set(22 posts)→
More from coding & agent
- Vercel Connect adds MCP Events in public beta for event-driven agents — cramforce · 2026-10-16
- PE operating partner shares real agent deployments: 59/60 accuracy at $250/month — alex_verem · 2026-10-16
- Dev: live voice + tool-calling + image gen now combine into a workable 'Her'-like experience — andreisavu · 2026-10-16
- Kinex: a harness that connects coding agents to any physical robot for scalable learning — daibond_alpha · 2026-10-16
- Developer rebuilds GitHub in 4 days: Keel makes agents declare intent before changing code — threepointone · 2026-10-16
- Stop using small models as judges with overloaded context, dev warns — xeophon · 2026-10-16